Market Regime Index Methodology — How a Regime Read Is Defined and Computed

What is a market regime read and what does it output?

A market regime read is a statement about the character of price movement over a trailing window — not its direction. We define it as four independent computations run side by side: a path-efficiency measure, a rolling Hurst exponent, an auction state derived from consecutive daily value areas, and a volatility compression state.

None of them is a prediction and none of them is combined into a composite index. Each one is a separate label attached to a bar, each one carries its own warm-up rule, and each one is allowed to return nothing. Inside the Confluence Engine, two of these reads become binary context flags on a zone so that hold-rate statistics can be split by regime after the fact rather than assumed in advance.

The output of a regime read is therefore a small set of labels with timestamps, not a score. That is a deliberate design choice and we state it as one.

How is the trend-versus-range input computed?

With Kaufman's efficiency ratio over a trailing window of fourteen bars: the absolute net move across the window divided by the summed absolute bar-to-bar path over the same window. A value near one means price walked in a straight line. A value near zero means it covered the same ground repeatedly.

The ratio uses only past bars and the bar being scored, so it is valid to consult at a historical decision time. That property is what allows regime-conditional statistics to mean anything.

The interesting part is not the formula, which is textbook, but the threshold. A fixed constant separating trend from range is the obvious implementation and it is wrong in a specific way: the same number means different things on different instruments. A ratio that is unusually high for a large-cap perpetual can be unremarkable on a thinner one. So the published regime label is derived from the percentile of the current ratio against that instrument's own strictly prior trailing distribution — the observation is ranked before being inserted into the history, never against itself — and the label reads trend when the ratio sits in the upper part of its own distribution.

We keep a fixed-threshold variant on the historical statistics path unchanged, on purpose, because the conditional statistics already measured against it would silently shift if we recalibrated underneath them.

What the Hurst input measures and where it fails

The second input is a rolling Hurst exponent estimated by classic rescaled range over a two-hundred-bar trailing window of log returns. For each of several sub-window sizes, the returns are split into non-overlapping chunks, each chunk is mean-centred, the range of its cumulative sum is divided by its own standard deviation, and the averages are regressed in log-log space. The slope is the exponent.

Readings above the upper band read as trending, below the lower band as mean-reverting, in between as neutral. Around one half is the random-walk reference.

Now the honest part, which we repeat wherever this appears: classic rescaled range is a biased estimator on windows this short, corrections exist and we do not apply them, and the reading is noisy near the window edges. Treat it as a gauge of which of three buckets a market is probably in, not as a fractal-dimension measurement.

There is a second limitation with real consequences. The estimator needs two hundred closed bars before it returns anything. Inside a seven-day computation window, minute and low-multiple-minute timeframes clear that easily, while hourly and coarser timeframes do not have two hundred bars at all. On those timeframes the Hurst-derived context flag is structurally unavailable — not tested and found irrelevant, simply unmeasured — and we label it that way rather than reading its absence as a result.

Balance or imbalance — the auction-state input

The third input asks a different question: is value being accepted two-sided, or is it migrating? It operates on completed daily volume profiles, comparing each day's value area against the previous day's.

The overlap between two value areas is computed as intersection over union — the Jaccard form. The alternative, dividing by the smaller value area, would credit a narrow area nested inside a much wider one as fully overlapping when it is really a slice of a larger balance region. When the overlap clears the threshold the state reads balance, and the bracket is the outer envelope of both value areas. Below it, value is migrating: up if the value-area midpoint rose, down if it fell, with the bracket being the current day's own value area.

The first day of any history emits no state. There is no principled neutral bracket without a second value area to compare against, so nothing is emitted — the same posture every other layer takes when evidence is insufficient. The state becomes knowable at the close of the later day, never before.

Which volatility input is used and why it is separate

The fourth input is a compression state per bar, and it is deliberately not folded into the other three. It combines a trailing percentile of average true range, a squeeze test — Bollinger bands contained entirely inside Keltner channels built on the same midline — and two narrow-bar flags: a lowest-true-range-of-seven bar and an inside bar.

A bar is marked compressed when the squeeze condition holds, or when the average-true-range percentile is low and one of the narrow-bar flags is set. Compression is context, not a trigger. It says the market is coiled, which is orthogonal to whether it is trending, and orthogonal again to whether daily value is being accepted.

Why we describe thresholds as tuned rather than universal

Because they are. The Hurst bands, the value-area overlap threshold and the compression percentile floor are design choices taken from published convention and our own reading, then held fixed. They are not laws of markets and we do not present them as such.

Two disciplines follow from admitting that. First, a threshold is never retuned to improve a historical statistic, because a statistic measured against a moving threshold measures the tuning rather than the market. Second, where a threshold is adaptive — the efficiency-ratio percentile — the adaptation is defined against strictly prior data, so an observation never influences the distribution it is ranked in.

This is the same standard applied to every layer described in confluence engine methodology and to the flow measurements in order flow report methodology.

How a regime read can be reproduced

Every input above is computable from public candle data alone, which is the point. The efficiency ratio needs closes, the Hurst estimate needs log returns, the auction state needs daily volume profiles, and compression needs true range plus two band constructions. None of it requires our recording infrastructure.

What our stack adds is the plumbing: computing each read once per bundle, stamping each with the moment it became knowable, and joining it to the zone statistics so a regime hypothesis can be falsified instead of asserted. That joining step is what NAIRO reads on air, and it is why a regime chip on the broadcast can read as nothing at all.

If you want to see the reads narrated live rather than described, the free delayed stream runs continuously, and the covered instruments are listed under markets. Background on the structural vocabulary used here is in market structure explained.